Do’s and Dont’s with
Likelihoods
Louis Lyons
IC and Oxford
CDF and CMS
Gran Sasso
Sept 2010
1
Topics
What it is
How it works: Resonance
Error estimates
Detailed example: Lifetime
Several Parameters
Extended maximum L
Do’s and Dont’s with L
****
2
4
How it works: Resonance
y~
Γ/2
(m-M0)2 + (Γ/2)2
m
Vary M
0
m
Vary Γ
5
6
7
Maximum likelihood error
Range of likely values of param μ from width of L or l dists.
If L(μ) is Gaussian, following definitions of σ are equivalent:
1) RMS of L(µ)
2) 1/√(-d2lnL / dµ2)
(Mnemonic)
3) ln(L(μ0±σ) = ln(L(μ0)) -1/2
If L(μ) is non-Gaussian, these are no longer the same
“Procedure 3) above still gives interval that contains the
true value of parameter μ with 68% probability”
Errors from 3) usually asymmetric, and asym errors are messy.
So choose param sensibly
e.g 1/p rather than p;
τ or λ
9
10
11
Several Parameters
12
13
14
15
DO’S AND DONT’S WITH L
• NORMALISATION FOR LIKELIHOOD
• JUST QUOTE UPPER LIMIT
• (ln L) = 0.5 RULE
• Lmax AND GOODNESS OF FIT
pU
•  L dp  0.90
pL
• BAYESIAN SMEARING OF L
• USE CORRECT L (PUNZI EFFECT)
16
NORMALISATION FOR LIKELIHOOD
 P(x |  ) dx
data
MUST be independent of 
param
e.g. Lifetime fit to t1, t2,………..tn
INCORRECT
P (t |  ) 
e - t /
Missing 1 / 
 
 too big
Reasonable 
t
17
2) QUOTING UPPER LIMIT
“We observed no significant signal, and our 90% conf
upper limit is …..”
Need to specify method e.g.
L
Chi-squared (data or theory error)
Frequentist (Central or upper limit)
Feldman-Cousins
Bayes with prior = const,
1/ 
1/ 

etc
“Show your L”
1) Not always practical
2) Not sufficient for frequentist methods
18
90% C.L. Upper Limits

x
x0
19
ΔlnL = -1/2 rule
If L(μ) is Gaussian, following definitions of σ are
equivalent:
1) RMS of L(µ)
2) 1/√(-d2L/dµ2)
3) ln(L(μ0±σ) = ln(L(μ0)) -1/2
If L(μ) is non-Gaussian, these are no longer the same
“Procedure 3) above still gives interval that contains the
true value of parameter μ with 68% probability”
Heinrich: CDF note 6438 (see CDF Statistics
Committee Web-page)
Barlow: Phystat05
20
COVERAGE
How often does quoted range for parameter include param’s true value?
N.B. Coverage is a property of METHOD, not of a particular exptl result
Coverage can vary with μ
Study coverage of different methods of Poisson parameter μ, from
observation of number of events n
100%
Nominal
value
Hope for:
C(  )

21
COVERAGE
If true for all  :
“correct coverage”
P<  for some  “undercoverage”
(this is serious !)
P>  for some  “overcoverage”
Conservative
Loss of rejection
power
22
Coverage : L approach (Not frequentist)
P(n,μ) = e-μμn/n!
-2 lnλ< 1
(Joel Heinrich CDF note 6438)
λ = P(n,μ)/P(n,μbest)
UNDERCOVERS
23
Frequentist central intervals, NEVER
undercovers
(Conservative at both ends)
24
Feldman-Cousins Unified intervals
Frequentist, so NEVER undercovers
25
Probability ordering
Frequentist, so NEVER undercovers
26
 2 = (n-µ)2/µ Δ 2 = 0.1
24.8% coverage?
NOT frequentist : Coverage = 0%  100%
27
Unbinned Lmax and Goodness of Fit?
Find params by maximising L
So larger L better than smaller L
So Lmax gives Goodness of Fit??
Bad
Good?
Great?
Monte Carlo distribution
of unbinned Lmax
Frequency
Lmax
28
Not necessarily:
L(data,params)
fixed vary
Contrast pdf(data,params)
pdf
L
param

vary fixed
e.g. p(λ) = λ exp(-λt)
data
Max at λ=1/t
Max at t = 0
L
p
t
λ
29
Example 1
Fit exponential to times t1, t2 ,t3 …….
[ Joel Heinrich, CDF 5639 ]
L = Π λ exp(-λti)
lnLmax = -N(1 + ln tav)
i.e. Depends only on AVERAGE t, but is
INDEPENDENT OF DISTRIBUTION OF t
(except for……..)
(Average t is a sufficient statistic)
Variation of Lmax in Monte Carlo is due to variations in samples’ average t , but
NOT TO BETTER OR WORSE FIT
pdf
Same average t
same Lmax
t
30
Example 2
1  cos 2 

d cos 
1  / 3
dN
L=
i
1   cos2 i
1  / 3
cos θ
pdf (and likelihood) depends only on cos2θi
Insensitive to sign of cosθi
So data can be in very bad agreement with expected distribution
e.g. all data with cosθ < 0
and Lmax does not know about it.
Example of general principle
31
Example 3
Fit to Gaussian with variable μ, fixed σ
1  x - 
pdf 
exp{- 

2  
 2
1
2
}
lnLmax = N(-0.5 ln2π – lnσ) – 0.5 Σ(xi – xav)2 /σ2
constant
~variance(x)
i.e. Lmax depends only on variance(x),
which is not relevant for fitting μ
(μest = xav)
Smaller than expected variance(x) results in larger Lmax
x
Worse fit, larger Lmax
x
Better fit, lower Lmax
32
Lmax and Goodness of Fit?
Conclusion:
L has sensible properties with respect to parameters
NOT with respect to data
Lmax within Monte Carlo peak is NECESSARY
not SUFFICIENT
(‘Necessary’ doesn’t mean that you have to do it!)
33
Binned data and Goodness of Fit using L-ratio
ni
μi
L=

Lbest

P n i (i )
i


P n i (i , best )
i

x
Pni (n i )
i
ln[L-ratio] = ln[L/Lbest]
large μi
-0.52
i.e. Goodness of Fit
μbest is independent of parameters of fit,
and so same parameter values from L or L-ratio
Baker and Cousins, NIM A221 (1984) 437
34
L and pdf
Example 1: Poisson
pdf = Probability density function for observing n, given μ
P(n;μ) = e -μ μn/n!
From this, construct L as
L(μ;n) = e -μ μn/n!
i.e. use same function of μ and n, but
. . . . . . . . . . pdf
for pdf, μ is fixed, but
for L, n is fixed
μ
L
n
N.B. P(n;μ) exists only at integer non-negative n
L(μ;n) exists only as continuous function of non-negative μ
35
Example 2
Lifetime distribution
pdf
p(t;λ) = λ e -λt
So
L(λ;t) = λ e –λt
(single observed t)
Here both t and λ are continuous
pdf maximises at t = 0
L maximises at λ = t
N.B. Functional form of P(t) and L(λ) are different
Fixed λ
Fixed t
L
p
t
λ
36
Example 3:
Gaussian
( x -  )2
pdf ( x ; ) 
exp {}
2
2

 2
1
( x - )2
L(; x ) 
exp {}
2
2
 2
1
N.B. In this case, same functional form for pdf and L
So if you consider just Gaussians, can be confused between pdf and L
So examples 1 and 2 are useful
37
Transformation properties of pdf and L
Lifetime example: dn/dt = λ e –λt
Change observable from t to y = √t
dn dn dt
-y 2

 2y e
dy
dt dy
So (a) pdf changes, BUT


(b)
dn
dn

t0
dt
dt 

t0
dy
dy
i.e. corresponding integrals of pdf are
INVARIANT
38
Now for Likelihood
When parameter changes from λ to τ = 1/λ
(a’) L does not change
dn/dt = 1/τ exp{-t/τ}
and so L(τ;t) = L(λ=1/τ;t)
because identical numbers occur in evaluations of the two L’s
BUT
(b’)
0
L
(;t ) d  
0

L
(;t ) d 
0
So it is NOT meaningful to integrate L
(However,………)
39
pdf(t;λ)
L(λ;t)
Value of
function
Changes when
observable is
transformed
INVARIANT wrt
transformation
of parameter
Integral of
function
INVARIANT wrt Changes when
transformation param is
of observable
transformed
Conclusion
Integrating L
Max prob
density not very not very
sensible
sensible
40
CONCLUSION:

pu
L dp   NOT recognised statistical procedure
pl
[Metric dependent:
τ range agrees with τpred
λ range inconsistent with 1/τpred ]
BUT
1) Could regard as “black box”
2) Make respectable by L
Bayes’ posterior
Posterior(λ) ~ L(λ)* Prior(λ)
[and Prior(λ) can be constant]
41
42
Getting L wrong: Punzi effect
Giovanni Punzi @ PHYSTAT2003
“Comments on L fits with variable resolution”
Separate two close signals, when resolution σ varies event
by event, and is different for 2 signals
e.g. 1) Signal 1 1+cos2θ
Signal 2
Isotropic
and different parts of detector give different σ
2) M (or τ)
Different numbers of tracks  different σM (or στ)
43
Events characterised by xi and σi
A events centred on x = 0
B events centred on x = 1
L(f)wrong = Π [f * G(xi,0,σi) + (1-f) * G(xi,1,σi)]
L(f)right = Π [f*p(xi,σi;A) + (1-f) * p(xi,σi;B)]
p(S,T) = p(S|T) * p(T)
p(xi,σi|A) = p(xi|σi,A) * p(σi|A)
= G(xi,0,σi) * p(σi|A)
So
L(f)right = Π[f * G(xi,0,σi) * p(σi|A) + (1-f) * G(xi,1,σi) * p(σi|B)]
If p(σ|A) = p(σ|B), Lright = Lwrong
but NOT otherwise
44
Giovanni’s Monte Carlo for
A : G(x,0, A)
B : G(x,1, B)
fA = 1/3
Lwrong
Lright
A
B
1.0
1 .0
0.336(3)
0.08
Same
1.0
1.1
0.374(4)
0.08
0. 333(0)
0
1.0
2.0
0.645(6)
0.12
0.333(0)
0
12
1.5 3
0.514(7)
0.14
0.335(2) 0.03
1.0
12
0.482(9)
0.09
0.333(0)
fA
f
fA
f
0
1) Lwrong OK for p(A)  p(B) , but otherwise BIASSED
2) Lright unbiassed, but Lwrong biassed (enormously)!
3) Lright gives smaller σf than Lwrong
45
Explanation of Punzi bias
σA = 1
σB = 2
A events with σ = 1
B events with σ = 2
x 
ACTUAL DISTRIBUTION
x
FITTING FUNCTION
[NA/NB variable, but same for A and B events]
Fit gives upward bias for NA/NB because (i) that is much better for A events; and
(ii) it does not hurt too much for B events
46
Another scenario for Punzi problem: PID
A
π
B
M
K
TOF
Originally:
Positions of peaks = constant
K-peak  π-peak at large momentum
σi variable, (σi)A = (σi)B
σi ~ constant, pK = pπ
COMMON FEATURE: Separation/Error = Constant
Where else??
MORAL: Beware of event-by-event variables whose pdf’s do not
appear in L
47
Avoiding Punzi Bias
BASIC RULE:
Write pdf for ALL observables, in terms of parameters
• Include p(σ|A) and p(σ|B) in fit
(But then, for example, particle identification may be determined more
by momentum distribution than by PID)
OR
• Fit each range of σi separately, and add (NA)i 
(NA)total, and similarly for B
Incorrect method using Lwrong uses weighted average
of (fA)j, assumed to be independent of j
Talk by Catastini at PHYSTAT05
48
Conclusions
How it works, and how to estimate errors
(ln L) = 0.5 rule and coverage
Several Parameters
Lmax and Goodness of Fit
Use correct L (Punzi effect)
49
Next time:
2
χ
and Goodness of Fit
Least squares best fit
Resume of straight line
Correlated errors
Errors in x and in y
Goodness of fit with χ2
Errors of first and second kind
Kinematic fitting
Toy example
THE paradox
50